STO-MP-IST-178 4 - 1 Ukrainian Conflict in Media: Two Approaches to Narrative Analysis Justina Mandravickaitė Tomas Krilavičius LITHUANIA [email protected]ABSTRACT Internet media is one of the most important tools to influence public opinion as well as reflect it. In this paper we analyze reflection of dynamics of Ukrainian conflict in BBC, RussiaToday, DayKiev and delfi.lt (main Lithuanian news portal). We apply two different approaches for the analysis: co-occurrence networks analysis to reflect change of rhetoric in four different media channels during conflict and sentiment-based storyline (syuzhet) analysis to monitor sentiment change in BBC from 2013 to 2014. We split conflict into three stages: beginning (Nov 21, 2013 – Jan 15, 2014), escalation (Jan 16, 2014 - Feb 17, 2014) and occupation of Crimea (Feb 18, 2014 – Feb 28, 2014). These approaches allow visual analysis of the conflict dynamics in media. Such application of Artificial Intelligence, Natural Language Processing and visualization techniques for big data allows better understanding of reflection of conflict dynamics and public mood on specific topics, automation of information analysis. 1.0 INTRODUCTION In narrative analysis it is essential to understand the sequence in which events occur [1, pp. 2-16]. Recent studies on narrative analysis focus on different types of events, e. g. linguistic-based narrative analysis of YouTube accounts [2]. It was used to describe common attributes of the narratives as well as to identify a list of shared thematic and linguistic characteristics. Study [3] reported how narratives were used by extremist organization. The researchers built a network of stories linked according to their semantic similarity. Moreover, it explored how narrative in social media shapes the vision of future by justification of present and explanation of past [4]. This study also showed how interconnected narratives provide intent and justification for different target audiences. Yet another study presented a case and usage of individual narrative [5]. According to this study, individual narratives help people to make sense of the world as well as shape and express their ideology, including the political functions. Usually narratives present events in chronological order, e.g., novels, personal retelling of experiences, etc. However, in news articles narratives differ from other types of narratives as they follow a complex time structure. In producing news stories, journalists present the events in “instalments”, thus an event that was introduced in the earlier parts of a story may be described in detail later, sometimes in multiple instances [6, pp. 147-174]. Therefore, events in news stories are usually presented in a non-chronological order, i.e., in the news stories storyline or syuzhet often does not match the chronological order of the events. In this paper we analyze reflection of dynamics of Ukrainian conflict in BBC, RussiaToday, DayKiev and delfi.lt (main Lithuanian news portal). We apply 2 different approaches for the analysis: word co-occurrence networks to reflect change of rhetoric in four different media channels during conflict as well as sentiment- based storyline (syuzhet) analysis to monitor sentiment change in BBC from 2013 to 2014.
17
Embed
Ukrainian Conflict in Media: Two Approaches to Narrative ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
STO-MP-IST-178 4 - 1
Ukrainian Conflict in Media: Two Approaches to Narrative Analysis
Ukrainian Conflict in Media: Two Approaches to Narrative Analysis
4 - 2 STO-MP-IST-178
2.0 DATA
We used 2 datasets for our research:
1. results of qualitative discourse analysis of media articles (BBC, RussiaToday, DayKiev and delfi.lt),
2. raw BBC articles.
The first dataset was prepared by the team of students mentored by scientists during the project “Research
Meadow / Mokslo pieva”. Media articles were collected from 4 different media sources: BBC, RussiaToday,
DayKiev and delfi.lt, from November 21, 2013 till February 28, 2014. This period covered 3 stages of the
Ukrainian conflict, see table 2-1 for some of the more important events for each of the Ukrainian conflict.
Table 2-1: Some of the more important events during 3 stages of Ukrainian crisis.
Conflict stage Events
1st (2013 11 21-2014 01 15)
● V. Yanukovych postpones the signing of an
association agreement with the EU.
● Y. Tymoshenko is prevented from leaving
Ukraine.
● Large-scale protests begin.
● Over 100,000 people participate in weekly
demonstrations in Kiev.
● First clashes with riot police and arrests occur.
● In early December, over 800,000 people participate
in demonstrations in Kiev.
2nd (2014 01 16-02 17)
January 16-23
● The Supreme Rada adopts anti-demonstration law.
● First deaths.
● Raid of government buildings in western Ukraine.
January 28-29
● The Prime Minister resigns.
● The laws directed against the protests are annulled.
3rd (2014 02 18-28)
● Clashes begin in earnest: casualties on both sides –
protesters as well police.
● Agreement between opposition and V.
Yanukovych (02 21)
● Y. Tymoshenko is liberated, Yanukovych leaves
the country (02 22).
February 27-28
● Russian forces occupy the Crimean parliament
building.
● Russian troops are deployed to airports and other
strategic sites.
Qualitative discourse analysis was performed for these media articles, identifying actors, qualities they were
attributed to, predicates of events as well as the qualities they were attributed to. This analysis was performed
for BBC, RussiaToday, DayKiev and delfi.lt articles separately. We used the dataset of these results for our
narrative analysis via word co-occurrence networks.
The second dataset consists of raw BBC articles, covering the same 3 stages of Ukrainian crisis as described
above. The articles were grouped according to the conflict stage they described, and no pre-processing of the
texts was performed. We used this dataset for our sentiment-based storyline analysis.
Ukrainian Conflict in Media: Two Approaches to Narrative Analysis
STO-MP-IST-178 4 - 3
3.0 METHODS
We used 2 methods for narrative analysis in terms of Ukrainian crisis. The first one is word co-occurrence
network analysis and the second one – sentiment-based storyline analysis. We used word co-occurrence
networks with dataset containing results of qualitative discourse analysis of media articles (BBC,
RussiaToday, DayKiev and delfi.lt), while for sentiment-based storyline analysis we used raw BBC articles.
In the following sections we describe these 2 methods in more detail.
3.1 Word Co-occurrence Network
Network analysis describes a group of methods that characterize relationships between units to be analysed,
i.e. nodes representing units to be analyzed and edges representing connections between nodes. Network
analysis is often used in the context of social networks, where nodes usually represent people, while edges
refer to friendships, affiliations or other types of relationships [7], [8, p. 10]. Though network analysis is
most often used in terms of relationships between people, it can also be applied to represent relationships
between words, e. g. [9], [10], [11]. Word co-occurrence network analysis is text length insensitive method
thus it is suitable even when dataset is made of short texts or just texts of uneven length [12]. For word co-
occurrence network analysis we used textnets1 package for R2.
Word co-occurrence network analysis was performed for our first dataset, made of results of qualitative
discourse analysis of media articles (BBC, RussiaToday, DayKiev and delfi.lt). The textual data was
tokenized and then lemmatized i.e., each word was replaced with its most basic syntactic form [11]. Then
part-of-speech tagger (integrated into textnets, based of language models of Universal Dependencies
project3) was applied in order to identify nouns and noun phrases which most likely describe content of the
text [12]. In the next step term frequency-inverse document frequency (TF-IDF) was characteristic was
calculated for each noun and noun phrase. This characteristic was calculated for each stage of the conflict
(see Table 2-1) separately. Then a bipartite word co-occurrence networks were generated for each stage of
the conflict. These bipartite networks linked media sources (BBC, RussiaToday, DayKiev and delfi.lt) based
on terms (nouns and noun phrases) in their messages. Finally the community detection algorithm was
applied in order to detect similarities in the selected media sources in terms of reporting on the Ukrainian
conflict [11], [12]. The results of word co-occurrence networks regarding Ukrainian conflict is presented in
Section 4.1.
3.2 Sentiment-Based Storyline Analysis
Sentiment analysis is important task of Natural Language Technologies and Artificial Intelligence, with
applications such as automated analysis of reviews and social media, monitoring of political issues, etc. [13].
Sentiment analysis assumes that emotions are important in communication among humans. Some recent
work in terms of analysis of literary texts has suggested that shifts in sentiment can be used for analysis of
plot development [14, pp. 73-110]. As the raw sentiment time series is very noisy, a smoothing filter is
usually applied for the raw data [15]. The result is a fairly smooth curve which represents generalized
trend of the sentiment development of the text. This method is based on controlled vocabulary and data
compression. For sentiment-based storyline analysis we used package syuzhet4 for R.
1 https://github.com/cbail/textnets 2 R is a software environment (free) for statistical computing and graphics, see: https://www.r-project.org/ 3 https://universaldependencies.org/ 4 https://cran.r-project.org/web/packages/syuzhet/